How to Cut Filler Words and Add Captions in Descript

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Cut Filler Words and Add Captions in Descript.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Cut Filler Words and Add Captions in Descript.

Descript cuts filler words and adds captions from the same transcript. Import the clip so Descript transcribes it, correct any misheard words, open AI Tools and run Remove filler words, review and delete the hits you want gone, then open the Captions panel and pick a style. Captions are built from the script, so every correction appears in them.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Order matters more than any setting. Fix the transcript before you remove fillers or style captions, because a misheard product name becomes a misspelt caption on every frame it appears. Filler detection only works on English transcripts and spends AI credits, and every cut is reversible: Descript hides removed footage instead of deleting it, so a bad cut costs a minute, not a re-shoot.

What you need before the first cut

A training clip is different from a marketing video. Nobody needs it to be slick, but new staff will watch it several times, pause it and copy what they see, so the words must be right and the pauses that show an action must survive. Get four things ready before you open the editor.

  • The raw file in a format Descript accepts. Descript imports MP4, M4V, MOV, WEBM and MKV video, and the upload size cap depends on the plan: 1 GB on Free, 10 GB on Hobbyist, 20 GB on Creator and 50 GB on Business. A phone recording of a ten-minute demonstration sits well inside all of those.
  • A list of your own words. Product names, supplier names, kit and jargon that a speech model will guess at. These go into the transcription glossary (covered below) before you record the next batch.
  • A plan with enough AI credits. Filler removal, Studio Sound, Shorten word gaps and Underlord all spend credits. Transcription itself is not metered on current plans; media minutes measure what you upload and record.
  • Someone who knows the task. The person who reviews the cut should be able to say whether a pause was dead air or the moment the presenter tied a knot.

If the bigger question is whether Descript is the right tool for your recordings at all, the breakdown of what each Descript plan gives in media hours and AI credits covers the costs. The rest of this page assumes you already have it open.

How do you get a training clip into Descript and transcribed?

Descript transcribes a clip automatically once you add it to a project's script and the upload finishes. Create a project, drag the file into the script editor, and wait for the transcript to appear before editing anything.

  1. Create a new project and give it the clip's working title (for example "Packing 04: wrapping a wardrobe").
  2. Drag the video file into the script editor. Transcription starts after the file has finished processing and uploading to the cloud.
  3. If the clip is not in English or mixes languages, set the language rather than leaving it on Auto. Descript lets you turn on "Always ask before transcribing" or set a default language, and a wrongly detected language is fixed with Re-transcribe from the file options menu.
  4. Check the speaker labels. If two people talk, make sure each line is assigned to the right person now, because separate caption layers per speaker depend on it later.

Time it takes depends on file length, audio quality and your connection. Descript's own guidance says accuracy can reach up to 95% with clear audio, which on a six-minute clip of roughly 900 spoken words still leaves around 45 words to check. Files longer than 15 hours fail to transcribe automatically, which is only a concern if you drop a full day's recording in as one file.

How do you correct the transcript without changing the video?

Descript corrects transcript text without touching the media when you use Correct instead of typing over words. Double-click a word, press C (or click Correct in the hover toolbar), type the fix, then choose Correct for that instance or Correct All for every instance in the project.

The difference matters. Selecting a word and typing over it in normal editing deletes the original word's audio along with its text, because deleting script text removes the media; Correct tells Descript the audio was fine and only the text was wrong, then re-aligns the words to the sound (you will see a dotted grey underline while it does). For a run of fixes, press Alt+C on Windows or Option+C on a Mac to enter Correct text mode, edit the script like a document, then press the same keys to leave.

A glossary stops the same mistakes in next week's clips

The transcription glossary is a word bank for names and jargon. Open the menu in the top-left corner of a project, choose Tools, then Transcription Glossary, and type one term per line. It holds up to 30 terms per Drive, works on English transcripts only and is shared by everyone on that Drive. Descript also adds a word automatically once you have corrected it at least three times across transcripts, a feature that is on by default.

Here is a filled-in glossary for an illustrative six-person removals firm that records short training clips for new packers:

Tri-wall
Bubble wrap (not "bubble rap")
Wardrobe carton
Piano skid
Sack truck
Furniture blanket
Stretch wrap
Moving dolly
Tail lift
Inventory sheet
Crate hire
Mattress bag

Twelve terms leaves eighteen slots for names of regular clients' buildings, staff and the van fleet. Once the list is in, new recordings should come back with fewer guesses, though you still read every line.

How do you remove filler words in Descript?

Descript removes filler words from the AI Tools panel: choose Remove filler words in the right sidebar, review the detected words with their timestamps, and pick how each one is handled.

Detected fillers such as "um" and "uh" show up in the script with a light blue underline. Descript's pricing page describes the tool as highlighting "um", "ah", "like", repeated words and more. The sidebar gives you four ways to treat each hit, and they behave very differently in a training clip:

OptionScriptAudio and videoUse it when
DeleteWord removedRemovedThe filler is dead weight and the cut sounds clean
Delete and replace with gapWord removedRemoved, with a gap clip that keeps the timingThe presenter's hands are mid-action and the picture must not jump
IgnoreWord struck through but still visibleRemoved from playbackYou want a visible record of the edit in case someone asks why a line changed
Remove from transcriptWord hidden from script, captions and exported transcriptsUnchangedThe sound is fine but you do not want "um" in the captions

Two settings make or break the result. Avoid harsh cuts tells Descript to skip fillers that cannot be removed without clipping the next word or leaving a jolt, after a quick audio analysis; keep it on for anything a customer or new starter will watch. And scope: the tool works on the whole composition by default. To clean only one section, duplicate that section into a new composition, run Remove filler words there, then paste the cleaned script back.

On cost, Descript caps Remove Filler Words at 30 AI credits per file or per application of the tool. A quick sum shows what that means: the Hobbyist plan's 400 monthly credits cover at least 13 filler passes even if every run hit the cap, and Studio Sound has the same 30-credit cap, so a clip that gets both costs at most 60 credits. Credits do not roll over; both balances reset when the subscription renews.

Shorten the pauses separately

Long silences are not filler words, so they need their own pass. In AI Tools choose Shorten word gaps, set the threshold (for example "more than" 1.5 seconds) and a target length such as 200 milliseconds, then step through results and click Shorten, or Shorten all once you trust the settings. This also spends AI credits. In a demonstration clip, review rather than shortening everything: a three-second pause while someone tapes a carton shut is the instruction.

Which filler words should stay in a training clip?

Training clips should keep any word that carries meaning or timing, even when Descript flags it. "Like" used as a comparison, "so" that introduces a result, and a repeated word that marks a correction ("the left, the left-hand panel") all read as filler to a detector.

A before-and-after from the illustrative removals firm's wardrobe clip shows the judgement calls. The raw transcript read:

So, um, the first thing, uh, you want to do is, like, check the
wardrobe carton is the right height. Um, you know, if it's too short
the, the rail will bend. So you, uh, you hook the rail in like this,
and, um... (3 seconds) ...then you close the top flaps.

After one review pass it read:

The first thing you want to do is check the wardrobe carton is the
right height. If it's too short the rail will bend. So you hook the
rail in like this... (3 seconds) ...then you close the top flaps.

What changed and why: the opening "So, um" went (Delete); "uh" and "you know" went; the repeated "the, the" went; but the second "So" stayed because it introduces the action, "like this" stayed because the presenter is pointing at the rail, and the three-second pause stayed because that is when the rail goes in. Descript flagged eleven items in this passage; the reviewer deleted eight and kept three. Expect a ratio like that on demonstration footage, and expect a higher deletion rate on talking-head clips where nothing happens on screen.

How do you add captions from the transcript in Descript?

Descript adds captions from the Captions panel in the right sidebar: apply them to one scene or all scenes, click a style, then adjust font, size and colours in the Properties panel.

  1. Open the right-hand sidebar and select the Captions panel.
  2. Use the toggle at the top to apply captions to the current scene or to all scenes. For a single-shot training clip, choose all scenes.
  3. Click a caption style. Search by name, filter by style, or sort by recently used; older styles sit under Legacy at the bottom of the list.
  4. Select the captions layer and open the Properties panel to set font, size, alignment, text and background colours, stroke, and how the active word is highlighted.
  5. If two people speak, add one captions layer per speaker and pick the speaker for each layer from the options slider beside the layer type, so each person's captions can sit in a different place.

Because captions are timed from the script, the earlier correction work is what makes them right. To switch styles later, select the existing captions layer first and then click the new style; clicking a style with nothing selected adds a second layer instead. Deleting a captions layer does not touch the script.

For training, readability beats animation. One-word styles that flash a single word at a time look lively on social feeds but are hard to follow when someone is also watching hands. A two-line style with a solid background, high contrast and the active word gently highlighted is easier on a warehouse tablet or a phone in a van cab.

Can Underlord do the clean-up for you?

Underlord, Descript's AI co-editor, can run the clean-up from a chat prompt, but it is in beta, spends AI credits on every prompt and tool, and needs checking like a new assistant's work.

Open it from the Underlord icon in the right-hand sidebar of any project. Descript lists caption styling, splitting videos by length, finding clips, translation, Studio Sound and audio balancing among the jobs it handles, and says tools such as Remove filler words can be requested through a prompt as well as run by hand. Each run charges for Underlord's reasoning plus the tool's work, with no fixed price per action; the Usage tab shows what each run cost. Descript also warns that Underlord may overpromise or make wrong assumptions, which is a fair description of any AI editor.

A prompt that works on a training clip names the limits, not only the goal:

Remove filler words (um, uh, you know) from the whole script, but keep
any pause longer than 2 seconds, keep "like this" where the speaker is
pointing at something, and turn on Studio Sound at a moderate intensity.
Then add captions in a two-line style with a dark background.
List every change you made.

An illustrative reply looks like this:

Done. I removed 47 filler words across the script and kept 6 pauses
longer than 2 seconds. I kept 3 instances of "like this". Studio Sound
is on at medium intensity. I added a captions layer using a two-line
style with a dark background, applied to all scenes.

What to fix in that output: it reports counts, not locations, so you still play the clip through at normal speed. Check the "kept" claims by searching the script for "like this". And check that the captions layer did not pick up a one-word style by default. If the result is wrong, use Revert under that response (covered below) rather than undoing step by step.

Save the routine as a skill

If you clean clips every week, turn the prompt into an Underlord skill. Open App settings (Ctrl+, on Windows or Cmd+, on a Mac), select the Skills tab, click New skill, choose Write skill manually, give it a unique name of up to 100 characters, add a short description, write the instructions in Markdown and save. In any project, type / in the Underlord chat box to pick it. Skills are private to the person who made them for now; sharing with a team or Drive is not available yet, so each editor saves their own copy.

Name: Training clip clean-up
Description: Fillers out, pauses kept, readable captions on.
Instructions:
- Remove um, uh, er and "you know" across the whole script.
- Never cut a pause over 2 seconds; list them instead.
- Keep words that point at actions ("like this", "here", "this one").
- Turn on Studio Sound at medium intensity.
- Add two-line captions with a dark background to all scenes.
- Finish with a list of every change and every item you skipped.

If you want help fitting clip clean-up into how your team records and signs off training, I offer one-to-one AI consultations for small businesses.

How do you undo a bad cut in Descript?

Descript undoes a bad cut in five ways, from the quickest to the widest: Undo, Restore removed media, un-ignoring a word, Revert on an Underlord response, and Version history for the whole project.

RouteHowReachWatch out for
UndoCtrl+Z (Cmd+Z on a Mac); redo with Ctrl+Shift+Z (Shift+Cmd+Z)One step per pressSlow after a bulk filler pass with dozens of cuts
Restore removed mediaSelect the area, right-click, Restore removed media; or Ctrl+Shift+Backspace (Shift+Cmd+Delete)Every cut inside the selectionOnly works where the script shows an edit boundary
Restore an ignored wordHover the struck-through text and choose RestoreOne ignored passageIgnored text is kept out of subtitles and web-link exports until restored
Revert (Underlord)Click Revert in the action bar under the Underlord responseEverything that response changedCheckpoints last for the current session only; a refresh or closing the project loses them
Version historyFile, then Version history, then Restore on an earlier versionThe whole projectRestores every composition and file, including a colleague's newer work

The safety comes from one design choice: Descript is non-destructive, so removed media stays in the project and is only hidden from the composition. That is why "Restore removed media" can bring back a section you cut an hour ago. Version history is the heavy tool. If two people edit the same project, restoring an old version can wipe the other person's changes, so agree a rule: restore only after telling whoever else is in the project.

A small habit saves most undo work: before running any bulk tool, duplicate the composition and name the copy "raw". If the bulk pass goes wrong after you have closed the project and lost the Revert checkpoint, the raw copy is still sitting there.

Worked example: a removals firm's six-minute packing clip

To see the whole sequence with numbers, follow one illustrative clip from raw file to published video. The firm records a supervisor demonstrating how to wrap and box a wardrobe for new packers; the files go on a staff learning page and a captioned cut goes on the firm's social accounts.

StepWhat happenedTimeAI credits
Import and transcribe6 min 40 s phone recording, about 930 words4 min waiting0 (transcription is not metered)
Correct transcript31 fixes, 9 of them "tri-wall" via Correct All; 4 terms added to the glossary9 min0
Remove filler words61 detected; 48 deleted, 9 ignored for the audit trail, 4 kept8 min reviewingUp to 30
Shorten word gaps14 gaps over 1.5 s found; 10 shortened to 300 ms, 4 kept as action pauses4 minCredits charged by job size
Studio SoundWarehouse echo reduced at medium intensity1 minUp to 30
CaptionsTwo-line style, dark background, all scenes3 min0
Review at normal speedTwo clipped words restored with Restore removed media7 min0
Export1080p MP4 for the learning page, SRT for the platform's player3 min0

The finished clip ran 5 minutes 52 seconds, 48 seconds shorter than the raw file, and took about 39 minutes of hands-on time including waiting. On the Hobbyist plan at $16 a month billed annually ($24 billed monthly), with 400 credits a month, this clip used at most 60 credits on the two capped tools plus a small charge for gap shortening. Eight clips a month at that rate would sit under roughly 500 credits in the worst case, which is where the Creator plan's 800 credits ($24 a month billed annually, $35 monthly) starts to make sense. Check the Usage tab after the first three clips rather than budgeting on worst cases; the real figure is usually lower than the caps.

Two things the numbers hide. First, the 9 minutes of correction work fell to 4 minutes on the next clip once the glossary held the firm's own terms. Second, the "4 kept" fillers were judgement calls only the supervisor could make; that review is the part to protect when the team is busy.

Which export format suits a training clip?

Descript exports a training clip as an MP4 through Export, top right, then Local export; a captions layer shows in the picture, while the embedded subtitle track is not burned in. Choose the combination by where the clip will play.

DestinationExportCaptionsWhy
Social feed (plays muted)Local export, Video tab, MP4, 1080pCaptions layer in the pictureViewers read before they unmute
Learning platform or YouTubeMP4 without a captions layer, plus Subtitles as SRT or VTTViewer switches them onAccessible, searchable and editable later
Quick internal shareDestination tab, Descript web linkCaptions layer if you added oneShare-page downloads carry no embedded subtitle track
Written procedureTranscript tab: .docx, .txt, .rtf, .md or .htmlNot applicableTurns the clip into a printable checklist
Further editing elsewhereTimeline export (Creator and Business)Depends on the editorKeeps the edit decisions instead of a flat video

In the Subtitles export, choose .srt or .vtt, decide whether to show speakers, and set Max characters per line and Max lines per card; two lines per card with short lines reads well on most players, so preview one export before you settle on a default. In the video export, resolutions run 480p, 720p, 1080p and 4K, but Free is limited to 720p and adds a Descript watermark in the bottom-right corner. Upgrading removes the watermark from future exports only, so anything exported on Free has to be exported again.

The double-caption problem

This is the mistake most teams hit once. A local MP4 export embeds a subtitle track automatically. If the composition also has a burned-in captions layer, a web player that shows embedded subtitles will display both, one over the other. Descript's own fix is to publish through the Descript web link destination, whose downloads carry no embedded subtitles, or to turn subtitles off in the player. In the removals firm example, the learning platform showed doubled captions on the first upload; re-exporting without the captions layer and uploading the SRT separately cured it.

Where do cleaned-up training clips still go wrong?

Training clips still go wrong in five predictable places after a Descript clean-up, and each one shows up only if someone watches at normal speed with the sound on.

  • Clipped words. A filler sitting tight against the next word takes its first syllable with it. "Avoid harsh cuts" reduces this; a listen-through catches the rest.
  • Jumps in the picture. Deleting a filler while hands are moving makes the image skip. Use Delete and replace with gap in demonstration sections.
  • Words that look fixed but are not. Typing over a word instead of using Correct removes the original audio for that word. If a sentence suddenly sounds clipped or silent, check whether a correction was made as an edit and restore the removed media.
  • Captions in the wrong language model. A clip transcribed on Auto that was really in another language produces nonsense captions; Re-transcribe with the right language before styling.
  • Old exports in circulation. Someone downloads the first version before corrections. Keep one place where the current file lives and date the file name.

Plan a library of clips like any other bit of admin. A simple log with one row per clip (title, recording date, reviewer, transcript checked, fillers reviewed, captions checked, export date, where it is published) stops a corrected version being overwritten by an old one. The team behind the illustrative removals firm keeps 40 clips this way and can see at a glance which ones need re-recording when a process changes.

If keeping that clip log and its review dates in a spreadsheet is the slow part, try The Complete Excel Course with AI Integration on Coding Liquids.

A filled-in sign-off checklist for one clip

Run this before any clip goes to new staff or out to customers. Here it is completed for the wardrobe clip:

CheckHow to verifyWardrobe clip
Product and kit names spelt rightSearch the script for each glossary termPass: "tri-wall" fixed in 9 places
No clipped wordsPlay at 1x with headphones2 found at 1:12 and 4:03, restored
Action pauses keptWatch every hand movementPass: 4 pauses kept
Captions readable on a phonePreview on a phone screenPass after switching to two-line style
Only one caption source per uploadUpload a test and play with subtitles onFixed: captions layer removed for the platform version
Right resolution, no watermarkCheck export settings and the first framesPass: 1080p on Hobbyist
Raw composition kept"Raw" copy visible in the projectPass

Which Descript plan does a training-clip routine need?

A business making a few training clips a month needs a paid Descript plan for watermark-free exports and regular filler removal; Hobbyist suits one editor and Creator suits heavier use or 4K. Business adds team-wide access to Brand Studio for a team that shares one look.

PlanPrice per person per monthMedia hoursAI creditsExportRemove Filler Words
Free$01 a month100, one time720p with watermarkLimited
Hobbyist$16 billed annually, $24 monthly10 a month400 a month1080p, no watermarkIncluded
Creator$24 billed annually, $35 monthly30 a month800 a month4K, no watermarkIncluded
Business$50 billed annually, $65 monthly40 a month1,500 a month4K, no watermarkIncluded

Prices and allowances are from Descript's pricing page as of November 2026; the lower figure is the annual-billing rate shown when the toggle is on annual. Dynamic captions are listed as unlimited on every plan, so captions themselves are never the reason to upgrade. The upgrade triggers are the watermark, 1080p or 4K, and how many credits your filler, Studio Sound and Underlord work burns each month. For the caption panel's full set of style controls, Descript's captions help page is the reference.

Three businesses, three clean-up settings

The same tools need different settings depending on what the clip shows. These illustrations come from the kinds of business this site's readers run.

A tutoring centre's onboarding clip for new tutors. Mostly a talking head explaining safeguarding steps and the booking system. Here filler removal can be aggressive (delete nearly everything flagged), word gaps can be shortened to 250 milliseconds across the board, and a one-line caption style keeps the screen clear for the slides behind the speaker. The risk is cutting so tightly that the speaker sounds rushed; listen to the first minute before applying Shorten all.

A landscaping company's "how we lay a patio base" clip. Outdoor sound, wind and a mechanical compactor in the background. Run Studio Sound first and lower its intensity if the waveform flattens (Descript notes very loud background noise can stop it separating speech from noise), then remove fillers with Delete and replace with gap so the footage of the compactor does not jump. Captions matter more than usual because half the audio is machinery.

An independent bookshop's staff clip on the returns process. Screen recording of the stock system with a voice-over. Fillers here often sit on top of clicks, so Ignore is the safer choice: the word leaves playback but stays visible as struck-through text, and the manager can see exactly what changed. Export the transcript as .docx afterwards and it doubles as the written procedure for the till.

For more on keeping captions, voice and intros consistent across a whole library, see how to keep AI-made videos on-brand. If the next job is cutting the same recordings into short social pieces, the walk-through on making captioned clips for social media covers that side, and if you are building training in several languages, making training videos in several languages with AI picks up where captions end. For teams who want AI to propose the cut itself, making a rough cut from an interview transcript shows how far that goes today.

Descript filler words and captions: quick answers

Does Descript remove filler words on the free plan?

Only in a limited way. Descript's pricing page lists Remove Filler Words as limited on Free and included on Hobbyist, Creator and Business. Detection works on English transcripts and spends AI credits, and Free gives 100 credits once rather than every month, so a team cleaning clips each week usually needs a paid plan.

Do Descript captions update when I correct the transcript?

Yes. Descript builds captions from the script, so a word fixed with Correct or Correct All shows up fixed in the captions layer and in any SRT or VTT file you export. Fix spelling before you style the captions, because changing a caption style never changes the words.

Will deleting a filler word cut the video as well?

Yes. Deleting a word in the script removes the matching audio and video, because the transcript is linked to the media. Choose Remove from transcript instead when you only want a word gone from the script, captions and transcript exports while the sound stays exactly as recorded.

Can I get my cuts back after running Remove filler words?

Yes. Press Ctrl+Z (Cmd+Z on a Mac) straight away, or select the cut area, right-click and choose Restore removed media, because Descript hides cut footage instead of erasing it. For a bigger rollback, File then Version history restores an earlier save, but it restores the whole project.

Should I burn captions in or upload an SRT file?

Burn captions in for social feeds where video plays muted, and upload an SRT or VTT file to a learning platform or YouTube, where viewers switch subtitles on and off. Avoid doing both on one upload: some players then show two sets of captions on top of each other.

Further reads

Show 4 more related tutorials

Sources: Descript pricing page; Descript help centre articles on filler words, edit like a doc, correcting transcripts, the transcription glossary, automatic transcription, captions, subtitle export, video export, embedded subtitles, restore removed media, deleting versus ignoring, version history, Underlord, reverting Underlord changes, Underlord skills, media minutes and AI credits, shortening word gaps, Studio Sound and supported file types (all read 22 November 2026).

Rather not work this out alone?

An AI implementation consultation is a one-to-one call with me, not an agency or a sales team. You leave knowing what AI you can use in your business and where to start.

Book a 1:1 call with me